① U.S. technology investment as a share of GDP has risen to approximately 4.9%, cloud vendors’ capital expenditure expectations for this year continue to be significantly revised upward, and AI infrastructure demand remains in expansion.
② Corporate operating conditions have not yet repeated the imbalances of the late 1990s, and AI supply chain profitability still has near-term support.
③ AI-related assets have accumulated approximately $27 trillion in value appreciation, and the market is betting on excess profits becoming long-term; either a slowdown in capital expenditure or diffusion of returns could trigger a revaluation.
From a macro perspective, this does not mean that current investment has already spiraled out of control.
Compared with the late 1990s, the current AI investment wave is not yet as broad as it was back then, and it has lasted for a shorter period, but its scale is rapidly approaching the level of that era. In the short term, as long as cloud providers continue to raise capital expenditure, there is still room for further improvement in orders, revenue and earnings expectations across the AI supply chain. Differences Between Today and the Late 1990s Goldman Sachs does not believe the current AI boom has fully replicated the bubble path of the late 1990s. At that time, while U.S. technology investment was overheating, corporate profit margins began to decline, corporate financing needs and leverage rose markedly, the current account deficit kept widening, and accelerating wage growth also pushed up unit labor costs. These factors ultimately weakened the foundation of economic expansion and laid the groundwork for the 2001 recession. The current situation is clearly different. First, U.S. corporate profits as a share of GDP remain close to a historical high, at about 14%. Productivity growth is relatively fast, but wage and unit labor cost increases are slower than in the late 1990s, meaning companies have not yet faced the kind of margin squeeze seen back then. Second, although free cash flow at cloud providers is being eroded by capital expenditure, the problem has not yet spread to the entire corporate sector. Goldman Sachs expects that by 2026, hyperscale cloud providers’ capital expenditure as a share of operating cash flow will rise to about 100%, reflecting that these companies are reinvesting almost all internal cash flow into AI infrastructure. However, overall capital expenditure pressure for S&P 500 constituents remains far below historical extreme levels. In addition, the U.S. current account deficit has not deteriorated as it did in the late 1990s, and the financial balance of the overall non-financial corporate sector has not worsened significantly. In other words, the AI investment boom has not yet triggered systemic credit expansion or macroeconomic financial imbalance. This is also the most critical difference between the current AI bull market and the 1999 technology stock frenzy: the rapid growth of corporate earnings has temporarily provided some real support for high valuations. The Market Has Run Ahead of Macro Realization Nevertheless, the market is still pricing the future returns of AI at a pace clearly faster than the actual realization of productivity dividends. Goldman Sachs estimates that since the release of ChatGPT at the end of November 2022, the combined market capitalization of AI-related listed companies and the valuations of private AI firms have increased by about $27 trillion, further rising from the $19 trillion recorded in November 2025. Even using more conservative estimation methods, the value the market assigns to AI remains enormous.If only “pure AI” assets such as semiconductor companies and private AI enterprises are counted, the value increment is approximately $14 trillion; if part of the gains from other AI-related companies is also included, the estimate rises to about $17 trillion.
By comparison, Goldman Sachs’ baseline estimate of the present value of additional capital income that AI can bring to businesses after boosting U.S. productivity is only $9 trillion. Even assuming U.S. companies can capture 50% of global AI gains, the valuation would only rise to $16 trillion; if a higher share of capital income and faster adoption speed are added, the estimate could rise to $18 trillion to $28 trillion. Only when multiple positive assumptions are combined and U.S. companies are assumed to capture a relatively large share of global AI gains could the potential value reach $43 trillion. The profit margins of AI winners may be more cyclical than the market imagines. The most noteworthy risk in the current AI market may not be a further rise in price-to-earnings ratios, but rather that peak earnings are being treated as the norm. Goldman Sachs pointed out that recent stock market gains have been driven more by upward revisions to earnings expectations than by further expansion of forward valuations. Forward P/E ratios have not risen significantly this year, but stock prices have continued to climb because the market keeps raising its earnings forecasts for 2026 and 2027. On the surface, this is healthier than pure valuation expansion. But if earnings growth itself is highly dependent on the capital expenditure cycle, the risk shifts from the valuation side to the earnings side. At present, profits in the semiconductor, server, networking equipment, and data center supply chains largely come from cloud providers’ concentrated construction of AI infrastructure. The investment cycle has not yet peaked, and in the short term orders and earnings forecasts may continue to be revised upward. But once capital expenditure growth slows, companies currently benefiting directly from the investment boom may be the first to face falling revenue growth, declining profit margins, and valuation reassessment. This is also why a low valuation does not necessarily mean cheap. If a company’s earnings are at a cyclical high driven by a capital expenditure boom, even if its forward P/E ratio appears low, it may not have sufficient margin of safety. In addition, the productivity gains brought by AI may not ultimately all flow to infrastructure suppliers. As technology spreads, enterprise customers, software companies, traditional industries, and consumers may all capture part of the gains. If the profit pool shifts to a broader range of industries, the excess profit share of current AI leaders may be diluted, and more intense sector rotation within the market will also occur. The non-AI economy is relatively weak, making the market more dependent on the AI narrative.Another layer of risk remains in the current macro backdrop: the AI boom is masking weakness in the non-AI parts of the U.S. economy.
Goldman Sachs notes that in the late 1990s, the U.S. economy experienced broad-based prosperity, with consumption, housing investment and non-tech investment all growing rapidly, and real domestic demand expanding at an annualized pace of nearly 6% in the final two years of the cycle. Today, by contrast, non-tech investment is generally weak, and consumption growth is not as strong as it was then. Over the past two years, U.S. real disposable income has grown at an annualized rate of about 1%, far below the roughly 5% to 6% level of the late 1990s. Although household wealth has risen and the savings rate has declined, these factors reflect stock market gains and the wealth effect rather than broad-based income growth. This means that AI investment is currently supporting the U.S. economy and corporate profits on the one hand, while making the market more dependent on a small number of industries and large companies on the other. Once AI capital expenditures, earnings expectations or the pace of technology commercialization fall short of expectations, the shock may not be confined to tech stocks but could transmit to the broader economy through the wealth effect, corporate investment and financial market sentiment. Goldman Sachs therefore believes that compared with 1999-2000, the risk of an extreme “pure valuation bubble” may be lower today, but the risk of an “earnings bubble” is rising. Volatility Rises, AI Bull Market Enters a More Divergent Phase Market structure has also begun to show characteristics similar to a late-stage bull market. Goldman Sachs observed that in recent months, implied volatility for individual U.S. stocks has risen notably, while implied correlation among stocks has fallen to a low level, meaning the market’s divergence between AI beneficiaries and those disrupted by AI is intensifying. The Philadelphia Semiconductor Index has risen by nearly 90% over the past two years, approaching the performance of the Nasdaq in the late 1990s. However, credit markets remain relatively calm for now. Increased AI-related financing has not significantly pushed up credit spreads, because investor risk appetite is strong and overall corporate leverage is markedly lower than in the late 1990s. Nevertheless, rising stock market volatility and widening divergence among individual stocks indicate that the market is no longer simply trading the single narrative of “AI growth” but is beginning to reassess who can truly capture AI profits, who may be disrupted by AI, and how long current capital expenditures can be sustained.
